pytorch/examples

Why don't we use MSE as a reconstruction loss for VAE ?

Open

#399 opened on Aug 7, 2018

View on GitHub
 (7 comments) (22 reactions) (0 assignees)Python (21,634 stars) (9,429 forks)batch import
good first issue

Description

Hi,

I am wondering if there is a theoretical reason for using BCE as a reconstruction loss for variation auto-encoders ? Can't we simply use MSE or norm-based reconstruction loss instead ?

Best Regards

Contributor guide

Tech stack
pythonpytorch
Domain
machine learning
Issue type
research
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
PyTorch basicsVariational Autoencoder concepts
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
60
Research direction
Investigate the theoretical reason for using BCE over MSE as reconstruction loss in VAEs. Refer to the original VAE paper (Kingma & Welling, 2013) which derives the ELBO with a Bernoulli likelihood. For continuous data, MSE corresponds to a Gaussian likelihood. Check the comments in the issue for discussion and the repository's existing VAE example to see current usage. Summarize the key point that the choice depends on the data distribution assumption.